llvm dev - Feb 2016 - WebKit B3 (was LLVM Weekly - #110, Feb 8th 2016)

LLVM Weekly - #110, Feb 8th 2016
===============================
If you prefer, you can read a HTML version of this email at
<http://llvmweekly.org/issue/110>.

Welcome to the one hundred and tenth issue of LLVM Weekly, a weekly newsletter
(published every Monday) covering developments in LLVM, Clang, and related
projects. LLVM Weekly is brought to you by [Alex
Bradbury](http://asbradbury.org). Subscribe to future issues at
<http://llvmweekly.org> and pass it on to anyone else you think may be
interested. Please send any tips or feedback to <asb at asbradbury.org>,
or
@llvmweekly or @asbradbury on Twitter.


## News and articles from around the web

Slides from the LLVM devroom at FOSDEM last weekend are [now available
online](http://llvm.org/devmtg/2016-01/). Unfortunately there was an issue
with the recording of the talks so videos will not be available.

JavaScriptCore's [FTL JIT](https://trac.webkit.org/wiki/FTLJIT) is moving
away
from using LLVM as its backend, towards [B3 (Bare Bones
Backend)](https://webkit.org/docs/b3/). This includes its own [SSA
IR](https://webkit.org/docs/b3/intermediate-representation.html),
optimisations, and instruction selection backend.

Source tarballs and binaries are [now
available](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094987.html)
for LLVM and Clang 3.8-RC2.

The Zurich LLVM Social [is coming up this
Thursday](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094759.html),
February 11th at 7pm.

Jeremy Bennett has written up a [comparison of the Clang and GCC command-line
flags](http://www.embecosm.com/2016/02/05/how-similar-are-gcc-and-llvm-the-user-perspective/).
The headline summary is that 397 work in both GCC and LLVM, 433 are LLVM-only
and 598 are GCC-only.

[vim-llvmcov](https://github.com/alepez/vim-llvmcov) has been released. It is
a vim plugin to show code coverage using the llvm-cov tool.


## On the mailing lists

* Mehdi Amini has posted an [RFC on floating point environment and rounding
mode handling in
LLVM](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094869.html). The
work started all the way back in 2014 and has a whole bunch of patches up for
review. Chandler Carruth has responded with a [detail description of his
concerns about the current
design](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094916.html),
and his proposed alternative seems to be getting a lot of positive feedback.

* Morten Brodersen has recently upgraded a number of applications from the old
JIT to the new MCJIT under LLVM 3.7.1 but has [found significant performance
regressions](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094908.html).
Some other respondents have seen similar issues, either in compilation time or
in reduced code quality in the generated code. Some of the thread participants
will be providing specific examples so they can be investigated. It's
possible
the issue is something as simple as a different default somewhere. Benoit
Belley noted [they saw regressions due to their frontend's use of allocas in
3.7](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094946.html).

* Lang Hames kicked off a long discussion about [error handling in LLVM
libraries](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094804.html).
Lang has implemented a new scheme and is seeking feedback on it. There's a
lot
of discussion that unfortunately I haven't had time to summarise properly.
If
error handling design interests you, do get stuck in.

* Adrian McCarthy has written up details on the [recent addition of minidump
support to
LLDB](http://lists.llvm.org/pipermail/lldb-dev/2016-February/009533.html).
Minidumps are the Windows equivalent of a core file.

* Juan Wajnerman is looking at adding support for multithreading to the
Crystal language, and has a [question about thread local
variables](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094736.html).
LLVM won't re-load the thread local address, which causes issues when a
thread
local variable is read in a coroutine running on one thread which is then
suspended and continued on a different thread. This is apparently a known
issue, covered by [PR19177](https://llvm.org/bugs/show_bug.cgi?id=19177).

* Steven Wu has posted an [RFC on embedding bitcode in object
files](http://lists.llvm.org/pipermail/llvm-dev/2016-February/094851.html).
The intent is to upstream support that already exists in Apple's fork.
Understandably some of the respondents asked how this relates to the .llvmbc
section that the Thin-LTO work is introducing. Steven indicates it's pretty
much the same, but for Mach-O rather than ELF and that he hopes to unify them
during the upstreaming.


## LLVM commits

* LLVM now has a memory SSA form. This isn't yet used by anything in-tree,
but
should form a very useful basis for a variety of analyses and transformations.
This patch has been baking for a long time, first being submitted for initial
feedback in April last year.
[r259595](http://reviews.llvm.org/rL259595).

* A new loop versioning loop-invariant code motion (LICM) pass was introduced.
This enables more opportunities for LICM by creating a new version of the loop
guarded by runtime checks to test for potential aliases that can't be
determined not to exist at compile-time.
[r259986](http://reviews.llvm.org/rL259986).

* LazyValueInfo gained an intersect operation on lattice values, which can be
used to exploit multiple sources of facts at once. The intent is to make
greater use of it, but already it is able to remove a half range-check when
performing jump-threading.
[r259461](http://reviews.llvm.org/rL259461).

* The SmallSet and SmallPtrSet templates will now error out if created with a
size greater than 32. [r259419](http://reviews.llvm.org/rL259419).

* The ability to emit errors from the backend for unsupported features has
been refactored, so BPF, WebAssembly, and AMDGPU backends can all share the
same implementation. [r259498](http://reviews.llvm.org/rL259498).

* A simple pass using LoopVersioning has been added, primarily for testing.
The new pass will fully disambiguate all may-aliasing memory accesses no
matter how many runtime checks are required.
[r259610](http://reviews.llvm.org/rL259610).

* The way bitsets are used to encode type information has now been documented.
[r259619](http://reviews.llvm.org/rL259619).

* You can now use the flag `-DLLVM_ENABLE_LTO` with CMake to build LLVM with
link-time optimisation. [r259766](http://reviews.llvm.org/rL259766).

* TableGen's AsmOperandClass gained the `IsOptional` field. Setting this to
1
means the operand is optional and the AsmParser will not emit an error if the
operand isn't present. [r259913](http://reviews.llvm.org/rL259913).

* There is now a scheduling model for the Exynos-M1.
[r259958](http://reviews.llvm.org/rL259958).


## Clang commits

* Clang now has builtins for the bitreverse intrinsic.
[r259671](http://reviews.llvm.org/rL259671).

* The option names for profile-guided optimisations with the cc1 driver have
been modified. [r259811](http://reviews.llvm.org/rL259811).


## Other project commits

* AddressSanitizer now supports iOS.
[r259451](http://reviews.llvm.org/rL259451).

* The current policy for using the new ELF LLD as a library has been
documented. [r259606](http://reviews.llvm.org/rL259606).

* Polly's new Sphinx documentation gained a guide on using Polly with Clang.
[r259767](http://reviews.llvm.org/rL259767).

.
>
> JavaScriptCore's [FTL JIT](https://trac.webkit.org/wiki/FTLJIT) is
moving
away
> from using LLVM as its backend, towards [B3 (Bare Bones
> Backend)](https://webkit.org/docs/b3/). This includes its own [SSA
> IR](https://webkit.org/docs/b3/intermediate-
<https://webkit.org/docs/b3/intermediate-representation.html>
representation.html
<https://webkit.org/docs/b3/intermediate-representation.html>),
> optimisations, and instruction selection backend.
In the end, what was the main motivation for creating a new IR?

Cheers,
Rafael
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160209/8d6c3e52/attachment.html>

> On Feb 9, 2016, at 9:55 AM, Rafael Espíndola via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> >
> > JavaScriptCore's [FTL JIT](https://trac.webkit.org/wiki/FTLJIT
<https://trac.webkit.org/wiki/FTLJIT>) is moving away
> > from using LLVM as its backend, towards [B3 (Bare Bones
> > Backend)](https://webkit.org/docs/b3/
<https://webkit.org/docs/b3/>). This includes its own [SSA
> > IR](https://webkit.org/docs/b3/intermediate-
<https://webkit.org/docs/b3/intermediate-representation.html>representation.html
<https://webkit.org/docs/b3/intermediate-representation.html>),
> > optimisations, and instruction selection backend.
> 
> In the end, what was the main motivation for creating a new IR?
> 
I can't speak to the motivation of the WebKit team. Those are outlined in
https://webkit.org/blog/5852/introducing-the-b3-jit-compiler/.
I'll give you my personal perspective on using LLVM for JITs, which may be
interesting to the LLVM community.

Most of the payoff for high level languages comes from the language-specific
optimizer. It was simpler for JavaScriptCore to perform loop optimization at
that level, so it doesn't even make use of LLVM's most powerful
optimizations, particularly SCEV based optimization. There is a relatively
small, finite amount of low-level optimization that is going to be important for
JavaScript benchmarks (most of InstCombine is not relevant).

SelectionDAG ISEL's compile time makes it a very poor choice for a JIT. We
never put the effort into making x86 FastISEL competitive for WebKit's
needs. The focus now is on Global ISEL, but that won't be ready for a while.

Even when LLVM's compile time problems are largely solved, and I believe
they can be, there will always be systemic compile time and memory overhead from
design decisions that achieve generality, flexibility, and layering. These are
software engineering tradeoffs.

It is possible to design an extremely lightweight SSA IR that works well in a
carefully controlled, fixed optimization pipeline. You then benefit from basic
SSA optimizations, which are not hard to write. You end up working with an IR of
arrays, where identifiers are indicies into the array. It's a different way
of writing passes, but very efficient. It's probably worth it for WebKit,
but not LLVM.

LLVM's patchpoints and stackmaps features are critical for managed runtimes.
However, directly supporting these features in a custom IR is simply more
convenient. It takes more time to make design changes to LLVM IR vs. a custom
IR. For example, LLVM does not yet support TBAA on calls, which would be very
useful for optimizating around patchpoints and runtime calls.

Prior to FTL, JavaScriptCore had no dependence on the LLVM project. Maintaining
a dependence on an external project naturally has integration overhead.

So, while LLVM is not the perfect JIT IR, it is very useful for JIT developers
who want a quick solution for low-level optimization and retargetable codegen.
WebKit FTL was a great example of using it to bootstrap a higher tier JIT.

To that end, I think it is important for LLVM to have a well-supported -Ojit
pipeline (compile fast) with the right set of passes for higher-level languages
(e.g. Tail Duplication).

-Andy
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160215/36427ef8/attachment.html>